Case study
Seekho
Edutainment OTT · $100M+ ARR · 100M downloads
How category-level personalization lifted notificationCTR +43% and doubled second-video starts
In six weeks, Pavo took Seekho's sticky notifications from broadcast logic to category-level personalization, lifting engagement across the full funnel, from click to second-video start, and deepening the retention a subscription business runs on.
- Monthly active users
- 25M
- Downloads
- 100M
- Annual recurring revenue
- $100M+
- +43%Notification CTR
- +94%Second-video initiation
- +27%70% video completion
The problem
Broadcast to everyone, personalized for no one
Seekho's notifications ran on a broadcast model: the same four daily pushes, at 7, 9, 11 AM and 1 PM, to everyone.
The candidates came from a handful of manual and query-driven sources:
- 01Content-manager requests.
- 02Event-driven manual picks: breaking news, trending topics, seasonal events.
- 03A query ranking content by total watch hours per business unit.
Two structural flaws were silently capping it:
- 01A self-reinforcing loop: the ranking counted views driven by earlier notifications as demand.
- 02Organic signal drowned out: notification-driven and organic views were blended, so the query couldn't tell content users sought out from content they were merely shown.
Content was organized into three business units (Awareness, Income, Skill) on the content side only; on the audience side there was no segmentation at all. The symptom was a click-through rate that had plateaued. But this wasn't a content-quality problem, Seekho's library has strong organic engagement, but a matching problem: the right content existed, it just wasn't reaching the right users. In a subscription business, every day a user doesn't find value is a day closer to churn.
“Getting a high CTR is easy, maybe by sending clickbaity notifications. That's why I want to track a more holistic metric: not just get the user on the platform, but let them consume the content.”
The approach
Map the system, fix the signal, ramp in phases
Pavo's principle: start with the broadest segmentation that could show lift in week one, validate it rigorously, build trust, then go deeper.
- 01
Compiled the system knowledge
- Connected to Seekho's warehouse and mapped every table in the notification pipeline into one operational model: content metadata, user profiles, video-play tables, MoEngage campaign data, and funnel tables.
- Reviewed the daily ranking queries that generated notification candidates.
- Surfaced the two structural flaws capping performance: the self-reinforcing notification loop, and organic signal drowned out by notification-driven views.
- 02
Designed the content-selection logic
- Took a mandate to power half the notification slots.
- One slot for proven organic performers: evergreen content up to 120 days old.
- One slot for fresh, trending content gaining organic momentum.
- Replaced raw view counts with a quality-weighted score across intent, completion depth, and second-video likelihood, plus category-diversity multipliers, saturation penalties, and organic-only counting.
- 03
Ramped up in phases
- Phase 1 (3 weeks): a 50/50 A/B matching each user to their primary interest group (Awareness, Income, Skill, or English), with two slots held as concurrent controls.
- Phase 2 (3 weeks): reclassified users into 16 category-level segments learned from watch behavior, rolled out to 100% across the 9 AM and 1 PM slots.
“The plan is very exhaustive: it covers all the cases any platform would want to cover, to move to signal-based, user-behavior-based notifications that improve the whole system, not just the CTR but the ROI overall.”
Segmentation depth
One audience became four groups, then sixteen
Broadcast treated every user the same. Pavo deepened the segmentation in two steps, proving each layer before going finer.
- 1
Broadcast
One push list for everyone, chosen by total watch hours.
- 4
Phase 1 · Interest groups
Matched to each user's primary interest, learned from watch behavior.
- Awareness
- Income
- Skill
- English
- 16Current
Phase 2 · Category segments
Finer clusters learned from watch behavior, the current layer and the floor of the next phase.
Each tier was validated before the next was built: interest groups proved the thesis, category segments deepened it.
The results
Lift across the full funnel, not just the click
Personalization moved every stage the notification touched, and Seekho held the gains stable week over week, the deeper daily habit a subscription business renews on.
- Notification CTR+43%
- Second-video initiation+94%
- 70% video completion+27%
Click-through rate
CTR climbed at every phase
The metric the old system was built around moved decisively, and Phase 1 ran against a concurrent broadcast control, so the lift is measured, not seasonal.
Notification CTR by phase · lift vs broadcast baseline
Broadcast (historical)
Phase 1 · Interest groups
Phase 2 · Category segments
Every value is a lift against the pre-Pavo broadcast baseline. The Phase-1 A/B held a concurrent broadcast control, which ran +18% on the same baseline against the personalized cells' +30%.
CTR was the surface metric even before Pavo, the number that made the problem visible. Even on the old system's own metric, the lift held.
Engagement depth
The fit compounded through every stage
From click to second-video start, every stage moved, and the deeper the stage, the larger the lift. A +43% CTR could be better thumbnails; a +94% lift in second-video starts cannot.
| Funnel stage | Broadcast | Phase 1 | Phase 2 |
|---|---|---|---|
| Click → 5s watched | Baseline | +6% | +18% |
| 40% video completion | Baseline | +20% | +23% |
| 70% video completion | Baseline | +22% | +27% |
| 2nd-video initiation | Baseline | +29% | +94% |
Every value is a lift against that stage's own broadcast baseline; the absolute rates are Seekho's and stay confidential. The pattern is monotonic across the funnel: Pavo found content users actually wanted to watch, and that fit compounded through every subsequent decision.
“We've seen a good jump in numbers and we've been able to maintain them in a stable way. The category-level bifurcation we did is the right spot.”
What's next
The 16-cluster layer is the floor, not the ceiling
The engagement-depth signals from Phase 2, especially the +94% on second-video starts, point to clear headroom above the current ladder. The 16-cluster layer is the floor of the next phase, not the ceiling.
| Phase | What it is |
|---|---|
| Phase 0 · Broadcast (historical) | Same content to all users |
| Phase 1 · Interest-group segments | 4 interest groups, group-level content; ranked by watch hours |
| Phase 2 · Category segments (current) | 16 category-based groups; ranked by deeper engagement metrics |
| Phase 3 · Per-user rules | Individual content from user intelligence (lifecycle, responsiveness, preference) and content intelligence (collaborative filtering, dedup, continuity) |
| Phase 4 · Per-user predictions | ML models rank content per user by predicted engagement; collaborative filtering becomes model-informed |
| Phase 5 · Self-learning | Adaptive exploration and weekly retraining; the system improves autonomously from feedback |
In closing
What made this possible
In six weeks Pavo mapped the notification system, found the structural flaws capping it, designed new quality-weighted selection logic, and ran a phased A/B program from interest groups to 16 category clusters, moving every stage of the funnel while holding CTR gains stable. The next target state is user-level, self-learning personalization above the cluster layer.
More case studies
Teams using Pavo to move the metrics that matter, in production, at scale.
- SeekhoEdutainment OTT · $100M+ ARR · 100M downloadsPavo built a pLTV model for Seekho to improve M3 retention by +8.23%+8.23%M3 retention lift · offline evalRead case study
- SeekhoEdutainment OTT · $100M+ ARR · 100M downloadsHow Pavo’s Segment-Adaptive Personalization Lifted Time Spent by 4.5%+4.5%Platform-wide time spentRead case study
- A high-traffic consumer platformRecommendation systems · ExperimentationBetter offline tests lead to more production wins−0.4 → +0.8Offline–online rank correlationRead case study
- DashverseShort-video entertainment · recommendation systemsOne improvement loop lifted Shortfree Autoplay activation by +3.1pp, live in production+3.1ppAutoplay activation · live in productionRead case study
